1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gemm/neonfp16arith-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/common.h>
16
17 #include <xnnpack/gemm.h>
18
19
xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const void* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 void* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(__fp16) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const __fp16* a0 = (const __fp16*) a;
42 __fp16* c0 = (__fp16*) c;
43
44 do {
45 float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
46
47 size_t k = kc;
48 while (k >= 4 * sizeof(__fp16)) {
49 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
50
51 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
52
53 #if XNN_ARCH_ARM64
54 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
55 #else
56 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
57
58 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
59 #endif
60 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
61
62 #if XNN_ARCH_ARM64
63 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
64 #else
65 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
66
67 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
68 #endif
69 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
70
71 #if XNN_ARCH_ARM64
72 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
73 #else
74 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
75
76 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
77 #endif
78 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
79
80 #if XNN_ARCH_ARM64
81 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
82 #else
83 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
84
85 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
86 #endif
87
88 k -= 4 * sizeof(__fp16);
89 }
90 if XNN_UNLIKELY(k != 0) {
91 do {
92 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
93
94 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
95
96 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
97
98 k -= sizeof(__fp16);
99 } while (k != 0);
100 }
101
102 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
103 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
104
105 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
106 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
107
108 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
109 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
110
111 if XNN_LIKELY(nc >= 8) {
112 vst1q_f16(c0, vacc0x01234567);
113 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
114
115 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
116
117 nc -= 8;
118 } else {
119 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
120 if (nc & 4) {
121 vst1_f16(c0, vacc0x0123); c0 += 4;
122
123 vacc0x0123 = vget_high_f16(vacc0x01234567);
124 }
125 if (nc & 2) {
126 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
127
128 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
129 }
130 if (nc & 1) {
131 vst1_lane_f16(c0, vacc0x0123, 0);
132 }
133
134 nc = 0;
135 }
136 } while (nc != 0);
137 }
138